Swift-YOLO: A Lightweight Architecture for Small Target Detection in Drone Optical Sensors

We propose Swift-YOLO, a lightweight framework for small target detection in resource-constrained drone optical sensor systems. Traditional deep learning detection methods exhibit inherent architectural limitations, including simplistic feature fusion strategies, fixed receptive fields, and redundant detection head parameters. These limitations are amplified in drone optical sensor imagery due to varying altitudes and scale inconsistencies, leading to insufficient accuracy and high computational overhead for small target detection. This paper designs three core modules. First, we design a lightweight hierarchical reconfigurable adaptive multi-scale feature pyramid network (LiteHRAMFPN) to extract reconfigurable, adaptive, and multi-scale features. This approach enhances the feature expression of small targets through adaptive weight fusion and multi-scale convolution. Second, the proposed method adopts a lightweight and efficient multi-scale convolution module (LEMSC) to achieve parallel cross-resolution feature learning using a channel grouping strategy. Third, we develop a lightweight, shared deformable convolution detection head (LSDCDH) to improve positioning accuracy through parameter sharing and deformable convolution. On the VisDrone2019 dataset, Swift-YOLO achieves an average precision of 18.2% and reduces the number of parameters by 38.7%. The generalization ability and practicality of the proposed method are verified on SIMD and DOTA datasets, proving its effectiveness for optical sensor-based aerial surveillance applications.